No Priors Ep. 141 | With Sunday Robotics Co-Founders Tony Zhao and Cheng Chi
Summary
Tony Zhao places robotics “in between the GPT moment and the ChatGPT moment”: the field appears to have a scalable recipe, but has not yet converted scale into a great consumer product. Classical sense-plan-act systems required bespoke interfaces for every task and environment; newer learning methods aim to scale data and models instead of repeatedly rebuilding task-specific systems. His bet is that robotics will follow other AI fields: more relevant data should improve usefulness.
Sunday’s central asset is an in-the-wild data engine with almost 10 million long-horizon trajectories, collected by more than 500 people. Cheng Chi’s UMI work replaced lab-bound teleoperation with a GoPro and gripper, while Diffusion Policy stabilized multimodal imitation, ALOHA made dexterous data collection more intuitive, and ACT/action chunking plus transformers helped make bimanual tasks scalable. A three-student collection produced 1,500 espresso-serving clips in roughly two weeks.
The apparent simplicity of glove-based collection hides full-stack execution difficulty. Sunday iterates the data device, robot controls, automated filtering and cleaning, calibration, training pipeline, and mechanical design together; the glove has gone from V0 to V5 with around 20 iterations per version. As Zhao recalls, the founders initially feared someone could “just take our glove,” but learned that “things are so much harder than we thought.”
Sunday rejects humanoid completeness when simplification can produce a useful robot faster and more cheaply. Its friendly, cartoon-like robot uses three fingers because people often use those fingers together for chores such as grasping handles or opening dishwashers. Perception lets the robot correct inexpensive, compliant, imprecise hardware, supporting a mechanically safe and compliant design while retaining sufficient accuracy.
The founders currently see imitation learning as more sample-efficient for manipulation, while reinforcement learning works well for locomotion. Ground contact is comparatively tractable to simulate, but reproducing a deforming hand, transparent cup, orange juice, reflections, and physical properties is extremely difficult and expensive. For manipulation, the behavior—put the hand in front of the cup and close with suitable force—can be easier to demonstrate than the world is to simulate.
Commercialization begins with a 2026 in-home beta, with shipping to the masses in 2027 or 2028 possible but explicitly dependent on reliability, capability, safety, and price. Prototypes currently cost $6,000-$20,000; at a few thousand units, Sunday expects material cost likely under $10,000 as costly low-volume cladding shifts to injection molding, implying a selling price around that level. Zhao envisions more than 1 billion home robots within a decade, while Cheng frames one possible future as the marginal cost of labor in homes approaching zero.
The founders’ rule for evaluating robotics videos is “make zero assumptions. No Priors.” Verify autonomy, then treat the exact object, person, environment, and demonstrated sequence as the proven capability—not evidence the robot can generalize. Sunday’s evidence includes long-horizon table cleanup, fragile glass handling, zero-shot trials across about six Airbnbs, espresso operation, and sock folding, though Cheng says the team shattered many glasses during development.
Deep dive
1. Robotics has a scaling recipe, but not yet its ChatGPT product
Zhao’s framing: robotics sits “in between the GPT moment and the ChatGPT moment.” Researchers increasingly agree on promising manipulation methods, but nobody yet knows what the jump from robotics’ equivalent of GPT-2 to GPT-3 will yield because the field has only recently obtained data at meaningful scale.
Classical robotics moved through human-designed sense-plan-act modules. Every task and environment demanded new interfaces—effectively “for every task, that means a paper”—so researchers and companies repeatedly discarded task-specific engineering instead of accumulating general capability.
Tony describes Diffusion Policy as stabilizing imitation learning when demonstrations contain multiple valid responses to the same observation. That made it possible for multiple, sometimes untrained operators to contribute without training diverging or the robot behaving strangely.
2. Better interfaces unlocked dexterity, then wild data unlocked scale
ALOHA made teleoperation feel more like playing a video game by greatly reducing the delay between human motion and robot response. Once the demonstrations became smooth and dextrous, transformers—which robotics had struggled to use after years of relying on 3-layer MLPs and ConvNets—began working well.
ACT and action chunking predict a trajectory rather than a single millisecond-scale action. Zhao’s intuition is biological: humans perceive, then move for a while without looking again, so chunking produces more consistent motion and better overall performance.
Chi’s escape from lab-bound teleoperation was UMI, using a 3D-printed gripper and GoPro to capture paired video and hand motion. Three students carried it into restaurants and gathered roughly 1,500 espresso-serving clips in two weeks, producing an unusually large dataset and an end-to-end policy that served drinks around Stanford in unseen locations.
The failure case was equally informative: the policy broke under direct sunlight because its collection period had been rainy. “In order for a robot to work in a sunny environment, it must have seen sunny environments” illustrates why wild-data breadth, not merely trajectory count, governs generalization.
3. Full-stack iteration turns collection scale into execution difficulty
Sunday grew from two founders clamping a robot to a desk in Chi’s apartment, to about eight people by late 2024, to around 30-40. Building a product rather than a demo required mechanical engineering, controls, software, AI, supply chain, and operations to optimize one system together. Because no general-purpose home robot exists, the right interfaces and standards are still unknown; the founders say this also makes outside partners difficult to use as their standard of “good” keeps changing.
Nearly 10 million wild trajectories now include navigation and extended tasks, not merely isolated cup pickups. More than 500 people using the gloves expose every failure mode; the V5 device followed around 20 iterations per version from V0 through V5, while automated calibration and fault detection protect data quality without requiring a human to inspect every clip.
4. A useful home robot wins by simplifying the humanoid
Sunday’s mission is to put a home robot in everyone’s home and remove chores that contribute little to what makes people “intrinsically human,” returning time for family, hobbies, and passions. If robots become cheap, safe, and capable, Zhao envisions more than 1 billion in homes within a decade.
Sarah Guo’s pushback—why not simply build a complete human form?—draws the core design answer: simplify wherever usefulness survives. Three fingers capture most benefits of grasping handles or opening dishwashers without multiplying actuator cost to separate fingers that ordinarily move together.
Industrial robots must be fast, stiff, and precise because they are “blindly following a trajectory.” Perception changes that constraint: cheap, compliant actuators may be mechanically inaccurate, but the AI system can correct hardware inaccuracies and deliver sufficient household accuracy while remaining “mechanically inherently safe and compliant.”
Commercialization remains conditional. A selected-user beta in 2026 will put real robots into homes; results determine whether shipping to the masses happens in 2027 or 2028. Prototypes cost $6,000-$20,000, but at a few thousand units Sunday expects material cost likely under $10,000, chiefly as CNC-machined, hand-painted cladding becomes injection-molded.
5. Manipulation favors imitation where simulation is hardest
The founders initially expected glove data to trail perfectly distribution-matched teleoperation. Instead, the form factor elicited more natural, dextrous behavior; after roughly 20 engineering iterations and full-cycle hardware/software work to convert human imagery into robot-like data, they no longer see a meaningful quality gap.
Reinforcement learning works well in environments that are easy to simulate, especially locomotion, where the relevant physical model is largely rigid-body dynamics and ground contact. Chi notes that a perfect simulator would make any task possible, so the practical question is which method gets there faster. Manipulation flips the equation: the behavior itself can be easier to capture, while accurately simulating transparent vessels, liquid color, reflections, deforming hands, contact, and force is extraordinarily expensive.
Data quality became more important with scale, not less. Hardware failures and uncontrolled wild behavior require continuous monitoring and repeatable cleaning; meanwhile, Sunday began serious research only about three months before the interview because “cute, fancy research ideas” developed under data scarcity might not scale into products.
Chi identifies two remaining challenges: finding a robust training recipe at scale and making hardware reliable. The learning team keeps pushing hardware toward its limits, so parts break; having mechanical and learning teams under one roof lets the company feed failures back into design quickly.
6. Demo credibility lives in autonomy, generalization, and sequence length
Zhao’s test begins with whether a demo is autonomous or teleoperated. If it shows one cup handed to one person, assume only that exact interaction works; viewers instinctively imagine different cups, people, laundry, and dishes, but should “only index on the things that are demonstrated strongly.”
Sequence length matters because every interaction introduces another failure probability. Sunday’s cleanup demo spans mobile manipulation, dumping food waste, loading dishes, and operating the dishwasher, and includes force-sensitive handling of two transparent wine glasses in one hand. Chi says, “We shattered a ton of glasses” during experimentation.
For generalization, the team booked around six Airbnbs and attempted tasks zero-shot: collect utensils into a caddy and load plates into dishwashers. The robot received no household-specific demonstrations yet handled reflective silverware, millimeter-sensitive grasps, and even a transparent table—coverage attributed to the diversity captured from more than 500 people in the dataset.
Espresso operation and sock folding test fine-grained force control. Conventional teleoperation leaves the operator’s hand effectively “numb,” allowing large forces to be applied without awareness; glove users can feel contact naturally. Sock folding also creates a force-closure loop where a stiff grip can apply effectively unlimited force without visible change, making the glove’s natural force feedback valuable for dexterity.